Qwen3-Next-80B-A3B-Instruct: 3B Active Parameters and RULER Results

What Qwen3-Next-80B-A3B-Instruct is

Qwen3-Next-80B-A3B-Instruct is Alibaba's efficiency-focused release: 80 billion total parameters with only 3 billion activated per token, open-weight and self-hostable, with a 256K token context window. It is roughly $0.09 per million input tokens through third-party providers.

Long-context results

Alibaba test on a 1M-token version of the RULER benchmark and publish accuracy at every length from 4K to 1,000K. Qwen3-Next-80B-A3B-Instruct averages 91.8%, against Qwen3-235B-A22B-Instruct-2507's 92.5%. At individual lengths: 99.0% at 8K against the 235B's 97.6%, 98.7% at 32K against 97.3%, 97.6% at 64K against 95.8%, and 96.0% at 128K against 93.9%. At a million tokens the 235B model scores 84.5% at a million tokens against this model's 80.3%. For comparison Qwen3-30B-A3B-Instruct-2507 averages 86.8% and falls to 72.8% at 1,000K.

The prompting requirement

Alibaba document that mathematics prompts should include the instruction "Please reason step by step, and put your final answer within \boxed{}". Published mathematics figures are produced with that phrasing.

LMArena

On LMArena's text leaderboard Qwen3 Next 80B A3b Instruct has 1399 Elo (rank 167); by category, 1444 Elo in Coding (rank 165), 1414 Elo in Math (rank 140), 1400 Elo in Expert (rank 183) and 1313 Elo in Creative Writing (rank 223).

Architecture

Qwen3-Next-80B-A3B is the first model in Alibaba's Qwen3-Next series. It replaces standard attention with a hybrid of Gated DeltaNet and gated attention, laid out as 12 repeats of three Gated DeltaNet layers followed by one gated-attention layer, each feeding a mixture-of-experts block: 48 layers in total with a hidden dimension of 2,048. Each MoE layer has 512 experts, with 10 routed and 1 shared expert active per token. Alibaba also describe zero-centred, weight-decayed layer norm for training stability, and multi-token prediction to improve pre-training and speed up inference. It was pre-trained on 15 trillion tokens; the native context is 262,144 tokens, extensible to 1,010,000. Alibaba report that the base model outperforms Qwen3-32B-Base on downstream tasks at 10% of the total training cost, with ten times the inference throughput for contexts over 32K tokens.

Alibaba’s comparison table

Against Qwen3-235B-A22B-Instruct-2507, Alibaba report MMLU-Pro 80.6 against 83.0, GPQA 72.9 against 77.5, AIME25 69.5 against 70.3, LiveCodeBench v6 56.6 against 51.8, Arena-Hard v2 82.7 against 79.2, WritingBench 87.3 against 85.2 and BFCL-v3 70.3 against 70.9. The model supports only non-thinking mode. Serving requires SGLang 0.5.2 or vLLM 0.10.2 or later; Alibaba's examples use tensor parallelism across 4 GPUs at a 256K context, with optional multi-token-prediction speculative decoding, which Hugging Face Transformers does not generally support.

More of Alibaba’s figures

Alibaba also report MultiPL-E 87.8, Aider-Polyglot 49.8, IFEval 87.6, TAU2-Retail 57.3, MultiIF 75.8 and PolyMATH 45.9, against Qwen3-235B-A22B-Instruct-2507's 87.9, 57.3, 88.7, 74.6, 77.5 and 50.2. Qwen3-32B in non-thinking mode records AIME25 20.2 and LiveCodeBench v6 29.1 in the same table, against this model's 69.5 and 56.6.

Tool calling

Alibaba point to Qwen-Agent for tool calling, which wraps tool-calling templates and parsers and accepts MCP server configuration files. If a server fails to start at the default 256K context, Alibaba suggest reducing it, for example to 32,768 tokens. Alibaba's example generates up to 16,384 new tokens.

Official Resources